docs/environment.md

Environment

This document records the exact environment used for this project.

Platform

• Compute: Google Colab Pro
• GPU: Tesla T4, 15.64 GB
• CUDA version: 12.1
• Python version: 3.12.13
• OS: Linux-6.6.113+-x86_64-with-glibc2.35

Key Library Versions

(See requirements.txt for the full pinned list)

• torch: 2.4.0+cu121
• torchaudio: 2.4.0+cu121
• transformers: 4.44.2

Storage

• Drive folder: /content/drive/MyDrive/deepfake_audio/
  • data/ — raw and preprocessed datasets
  • checkpoints/ — saved model weights
  • logs/ — local log files

Experiment Tracking

• Wandb project: deepfake-audio-detection
• Wandb entity: sara-jaffrani17-dlp

Reproducibility

• Random seed: 42
• Sample rate: 16 kHz mono throughout
• Window length: 4 seconds with 50% overlap
• Padding for short clips: zero-pad to 4 seconds

Phase 1 Verification

Phase 1 smoke test: 2026-04-28 All checks passed: GPU detected, Drive mounted, Wav2Vec 2.0 loaded, forward pass succeeded, wandb run logged.